Spaces:
Running
on
T4
Running
on
T4
add code
Browse files- app.py +553 -0
- requirements.txt +7 -0
- show.py +28 -0
app.py
ADDED
@@ -0,0 +1,553 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# --------------------------------------------------------
|
2 |
+
# PersonalizeSAM -- Personalize Segment Anything Model with One Shot
|
3 |
+
# Licensed under The MIT License [see LICENSE for details]
|
4 |
+
# --------------------------------------------------------
|
5 |
+
from PIL import Image
|
6 |
+
import torch
|
7 |
+
import torch.nn as nn
|
8 |
+
import gradio as gr
|
9 |
+
import numpy as np
|
10 |
+
from torch.nn import functional as F
|
11 |
+
|
12 |
+
from show import *
|
13 |
+
from per_segment_anything import sam_model_registry, SamPredictor
|
14 |
+
|
15 |
+
|
16 |
+
class ImageMask(gr.components.Image):
|
17 |
+
"""
|
18 |
+
Sets: source="canvas", tool="sketch"
|
19 |
+
"""
|
20 |
+
|
21 |
+
is_template = True
|
22 |
+
|
23 |
+
def __init__(self, **kwargs):
|
24 |
+
super().__init__(source="upload", tool="sketch", interactive=True, **kwargs)
|
25 |
+
|
26 |
+
def preprocess(self, x):
|
27 |
+
return super().preprocess(x)
|
28 |
+
|
29 |
+
|
30 |
+
class Mask_Weights(nn.Module):
|
31 |
+
def __init__(self):
|
32 |
+
super().__init__()
|
33 |
+
self.weights = nn.Parameter(torch.ones(2, 1, requires_grad=True) / 3)
|
34 |
+
|
35 |
+
|
36 |
+
def point_selection(mask_sim, topk=1):
|
37 |
+
# Top-1 point selection
|
38 |
+
w, h = mask_sim.shape
|
39 |
+
topk_xy = mask_sim.flatten(0).topk(topk)[1]
|
40 |
+
topk_x = (topk_xy // h).unsqueeze(0)
|
41 |
+
topk_y = (topk_xy - topk_x * h)
|
42 |
+
topk_xy = torch.cat((topk_y, topk_x), dim=0).permute(1, 0)
|
43 |
+
topk_label = np.array([1] * topk)
|
44 |
+
topk_xy = topk_xy.cpu().numpy()
|
45 |
+
|
46 |
+
# Top-last point selection
|
47 |
+
last_xy = mask_sim.flatten(0).topk(topk, largest=False)[1]
|
48 |
+
last_x = (last_xy // h).unsqueeze(0)
|
49 |
+
last_y = (last_xy - last_x * h)
|
50 |
+
last_xy = torch.cat((last_y, last_x), dim=0).permute(1, 0)
|
51 |
+
last_label = np.array([0] * topk)
|
52 |
+
last_xy = last_xy.cpu().numpy()
|
53 |
+
|
54 |
+
return topk_xy, topk_label, last_xy, last_label
|
55 |
+
|
56 |
+
|
57 |
+
def calculate_dice_loss(inputs, targets, num_masks = 1):
|
58 |
+
"""
|
59 |
+
Compute the DICE loss, similar to generalized IOU for masks
|
60 |
+
Args:
|
61 |
+
inputs: A float tensor of arbitrary shape.
|
62 |
+
The predictions for each example.
|
63 |
+
targets: A float tensor with the same shape as inputs. Stores the binary
|
64 |
+
classification label for each element in inputs
|
65 |
+
(0 for the negative class and 1 for the positive class).
|
66 |
+
"""
|
67 |
+
inputs = inputs.sigmoid()
|
68 |
+
inputs = inputs.flatten(1)
|
69 |
+
numerator = 2 * (inputs * targets).sum(-1)
|
70 |
+
denominator = inputs.sum(-1) + targets.sum(-1)
|
71 |
+
loss = 1 - (numerator + 1) / (denominator + 1)
|
72 |
+
return loss.sum() / num_masks
|
73 |
+
|
74 |
+
|
75 |
+
def calculate_sigmoid_focal_loss(inputs, targets, num_masks = 1, alpha: float = 0.25, gamma: float = 2):
|
76 |
+
"""
|
77 |
+
Loss used in RetinaNet for dense detection: https://arxiv.org/abs/1708.02002.
|
78 |
+
Args:
|
79 |
+
inputs: A float tensor of arbitrary shape.
|
80 |
+
The predictions for each example.
|
81 |
+
targets: A float tensor with the same shape as inputs. Stores the binary
|
82 |
+
classification label for each element in inputs
|
83 |
+
(0 for the negative class and 1 for the positive class).
|
84 |
+
alpha: (optional) Weighting factor in range (0,1) to balance
|
85 |
+
positive vs negative examples. Default = -1 (no weighting).
|
86 |
+
gamma: Exponent of the modulating factor (1 - p_t) to
|
87 |
+
balance easy vs hard examples.
|
88 |
+
Returns:
|
89 |
+
Loss tensor
|
90 |
+
"""
|
91 |
+
prob = inputs.sigmoid()
|
92 |
+
ce_loss = F.binary_cross_entropy_with_logits(inputs, targets, reduction="none")
|
93 |
+
p_t = prob * targets + (1 - prob) * (1 - targets)
|
94 |
+
loss = ce_loss * ((1 - p_t) ** gamma)
|
95 |
+
|
96 |
+
if alpha >= 0:
|
97 |
+
alpha_t = alpha * targets + (1 - alpha) * (1 - targets)
|
98 |
+
loss = alpha_t * loss
|
99 |
+
|
100 |
+
return loss.mean(1).sum() / num_masks
|
101 |
+
|
102 |
+
|
103 |
+
def inference(ic_image, ic_mask, image1, image2):
|
104 |
+
# in context image and mask
|
105 |
+
ic_image = np.array(ic_image.convert("RGB"))
|
106 |
+
ic_mask = np.array(ic_mask.convert("RGB"))
|
107 |
+
|
108 |
+
sam_type, sam_ckpt = 'vit_h', 'sam_vit_h_4b8939.pth'
|
109 |
+
sam = sam_model_registry[sam_type](checkpoint=sam_ckpt).cuda()
|
110 |
+
# sam = sam_model_registry[sam_type](checkpoint=sam_ckpt)
|
111 |
+
predictor = SamPredictor(sam)
|
112 |
+
|
113 |
+
# Image features encoding
|
114 |
+
ref_mask = predictor.set_image(ic_image, ic_mask)
|
115 |
+
ref_feat = predictor.features.squeeze().permute(1, 2, 0)
|
116 |
+
|
117 |
+
ref_mask = F.interpolate(ref_mask, size=ref_feat.shape[0: 2], mode="bilinear")
|
118 |
+
ref_mask = ref_mask.squeeze()[0]
|
119 |
+
|
120 |
+
# Target feature extraction
|
121 |
+
print("======> Obtain Location Prior" )
|
122 |
+
target_feat = ref_feat[ref_mask > 0]
|
123 |
+
target_embedding = target_feat.mean(0).unsqueeze(0)
|
124 |
+
target_feat = target_embedding / target_embedding.norm(dim=-1, keepdim=True)
|
125 |
+
target_embedding = target_embedding.unsqueeze(0)
|
126 |
+
|
127 |
+
output_image = []
|
128 |
+
|
129 |
+
for test_image in [image1, image2]:
|
130 |
+
print("======> Testing Image" )
|
131 |
+
test_image = np.array(test_image.convert("RGB"))
|
132 |
+
|
133 |
+
# Image feature encoding
|
134 |
+
predictor.set_image(test_image)
|
135 |
+
test_feat = predictor.features.squeeze()
|
136 |
+
|
137 |
+
# Cosine similarity
|
138 |
+
C, h, w = test_feat.shape
|
139 |
+
test_feat = test_feat / test_feat.norm(dim=0, keepdim=True)
|
140 |
+
test_feat = test_feat.reshape(C, h * w)
|
141 |
+
sim = target_feat @ test_feat
|
142 |
+
|
143 |
+
sim = sim.reshape(1, 1, h, w)
|
144 |
+
sim = F.interpolate(sim, scale_factor=4, mode="bilinear")
|
145 |
+
sim = predictor.model.postprocess_masks(
|
146 |
+
sim,
|
147 |
+
input_size=predictor.input_size,
|
148 |
+
original_size=predictor.original_size).squeeze()
|
149 |
+
|
150 |
+
# Positive-negative location prior
|
151 |
+
topk_xy_i, topk_label_i, last_xy_i, last_label_i = point_selection(sim, topk=1)
|
152 |
+
topk_xy = np.concatenate([topk_xy_i, last_xy_i], axis=0)
|
153 |
+
topk_label = np.concatenate([topk_label_i, last_label_i], axis=0)
|
154 |
+
|
155 |
+
# Obtain the target guidance for cross-attention layers
|
156 |
+
sim = (sim - sim.mean()) / torch.std(sim)
|
157 |
+
sim = F.interpolate(sim.unsqueeze(0).unsqueeze(0), size=(64, 64), mode="bilinear")
|
158 |
+
attn_sim = sim.sigmoid_().unsqueeze(0).flatten(3)
|
159 |
+
|
160 |
+
# First-step prediction
|
161 |
+
masks, scores, logits, _ = predictor.predict(
|
162 |
+
point_coords=topk_xy,
|
163 |
+
point_labels=topk_label,
|
164 |
+
multimask_output=False,
|
165 |
+
attn_sim=attn_sim, # Target-guided Attention
|
166 |
+
target_embedding=target_embedding # Target-semantic Prompting
|
167 |
+
)
|
168 |
+
best_idx = 0
|
169 |
+
|
170 |
+
# Cascaded Post-refinement-1
|
171 |
+
masks, scores, logits, _ = predictor.predict(
|
172 |
+
point_coords=topk_xy,
|
173 |
+
point_labels=topk_label,
|
174 |
+
mask_input=logits[best_idx: best_idx + 1, :, :],
|
175 |
+
multimask_output=True)
|
176 |
+
best_idx = np.argmax(scores)
|
177 |
+
|
178 |
+
# Cascaded Post-refinement-2
|
179 |
+
y, x = np.nonzero(masks[best_idx])
|
180 |
+
x_min = x.min()
|
181 |
+
x_max = x.max()
|
182 |
+
y_min = y.min()
|
183 |
+
y_max = y.max()
|
184 |
+
input_box = np.array([x_min, y_min, x_max, y_max])
|
185 |
+
masks, scores, logits, _ = predictor.predict(
|
186 |
+
point_coords=topk_xy,
|
187 |
+
point_labels=topk_label,
|
188 |
+
box=input_box[None, :],
|
189 |
+
mask_input=logits[best_idx: best_idx + 1, :, :],
|
190 |
+
multimask_output=True)
|
191 |
+
best_idx = np.argmax(scores)
|
192 |
+
|
193 |
+
final_mask = masks[best_idx]
|
194 |
+
mask_colors = np.zeros((final_mask.shape[0], final_mask.shape[1], 3), dtype=np.uint8)
|
195 |
+
mask_colors[final_mask, :] = np.array([[128, 0, 0]])
|
196 |
+
output_image.append(Image.fromarray((mask_colors * 0.6 + test_image * 0.4).astype('uint8'), 'RGB'))
|
197 |
+
|
198 |
+
return output_image[0].resize((224, 224)), output_image[1].resize((224, 224))
|
199 |
+
|
200 |
+
|
201 |
+
def inference_scribble(image, image1, image2):
|
202 |
+
# in context image and mask
|
203 |
+
ic_image = image["image"]
|
204 |
+
ic_mask = image["mask"]
|
205 |
+
ic_image = np.array(ic_image.convert("RGB"))
|
206 |
+
ic_mask = np.array(ic_mask.convert("RGB"))
|
207 |
+
|
208 |
+
sam_type, sam_ckpt = 'vit_h', 'sam_vit_h_4b8939.pth'
|
209 |
+
sam = sam_model_registry[sam_type](checkpoint=sam_ckpt).cuda()
|
210 |
+
# sam = sam_model_registry[sam_type](checkpoint=sam_ckpt)
|
211 |
+
predictor = SamPredictor(sam)
|
212 |
+
|
213 |
+
# Image features encoding
|
214 |
+
ref_mask = predictor.set_image(ic_image, ic_mask)
|
215 |
+
ref_feat = predictor.features.squeeze().permute(1, 2, 0)
|
216 |
+
|
217 |
+
ref_mask = F.interpolate(ref_mask, size=ref_feat.shape[0: 2], mode="bilinear")
|
218 |
+
ref_mask = ref_mask.squeeze()[0]
|
219 |
+
|
220 |
+
# Target feature extraction
|
221 |
+
print("======> Obtain Location Prior" )
|
222 |
+
target_feat = ref_feat[ref_mask > 0]
|
223 |
+
target_embedding = target_feat.mean(0).unsqueeze(0)
|
224 |
+
target_feat = target_embedding / target_embedding.norm(dim=-1, keepdim=True)
|
225 |
+
target_embedding = target_embedding.unsqueeze(0)
|
226 |
+
|
227 |
+
output_image = []
|
228 |
+
|
229 |
+
for test_image in [image1, image2]:
|
230 |
+
print("======> Testing Image" )
|
231 |
+
test_image = np.array(test_image.convert("RGB"))
|
232 |
+
|
233 |
+
# Image feature encoding
|
234 |
+
predictor.set_image(test_image)
|
235 |
+
test_feat = predictor.features.squeeze()
|
236 |
+
|
237 |
+
# Cosine similarity
|
238 |
+
C, h, w = test_feat.shape
|
239 |
+
test_feat = test_feat / test_feat.norm(dim=0, keepdim=True)
|
240 |
+
test_feat = test_feat.reshape(C, h * w)
|
241 |
+
sim = target_feat @ test_feat
|
242 |
+
|
243 |
+
sim = sim.reshape(1, 1, h, w)
|
244 |
+
sim = F.interpolate(sim, scale_factor=4, mode="bilinear")
|
245 |
+
sim = predictor.model.postprocess_masks(
|
246 |
+
sim,
|
247 |
+
input_size=predictor.input_size,
|
248 |
+
original_size=predictor.original_size).squeeze()
|
249 |
+
|
250 |
+
# Positive-negative location prior
|
251 |
+
topk_xy_i, topk_label_i, last_xy_i, last_label_i = point_selection(sim, topk=1)
|
252 |
+
topk_xy = np.concatenate([topk_xy_i, last_xy_i], axis=0)
|
253 |
+
topk_label = np.concatenate([topk_label_i, last_label_i], axis=0)
|
254 |
+
|
255 |
+
# Obtain the target guidance for cross-attention layers
|
256 |
+
sim = (sim - sim.mean()) / torch.std(sim)
|
257 |
+
sim = F.interpolate(sim.unsqueeze(0).unsqueeze(0), size=(64, 64), mode="bilinear")
|
258 |
+
attn_sim = sim.sigmoid_().unsqueeze(0).flatten(3)
|
259 |
+
|
260 |
+
# First-step prediction
|
261 |
+
masks, scores, logits, _ = predictor.predict(
|
262 |
+
point_coords=topk_xy,
|
263 |
+
point_labels=topk_label,
|
264 |
+
multimask_output=False,
|
265 |
+
attn_sim=attn_sim, # Target-guided Attention
|
266 |
+
target_embedding=target_embedding # Target-semantic Prompting
|
267 |
+
)
|
268 |
+
best_idx = 0
|
269 |
+
|
270 |
+
# Cascaded Post-refinement-1
|
271 |
+
masks, scores, logits, _ = predictor.predict(
|
272 |
+
point_coords=topk_xy,
|
273 |
+
point_labels=topk_label,
|
274 |
+
mask_input=logits[best_idx: best_idx + 1, :, :],
|
275 |
+
multimask_output=True)
|
276 |
+
best_idx = np.argmax(scores)
|
277 |
+
|
278 |
+
# Cascaded Post-refinement-2
|
279 |
+
y, x = np.nonzero(masks[best_idx])
|
280 |
+
x_min = x.min()
|
281 |
+
x_max = x.max()
|
282 |
+
y_min = y.min()
|
283 |
+
y_max = y.max()
|
284 |
+
input_box = np.array([x_min, y_min, x_max, y_max])
|
285 |
+
masks, scores, logits, _ = predictor.predict(
|
286 |
+
point_coords=topk_xy,
|
287 |
+
point_labels=topk_label,
|
288 |
+
box=input_box[None, :],
|
289 |
+
mask_input=logits[best_idx: best_idx + 1, :, :],
|
290 |
+
multimask_output=True)
|
291 |
+
best_idx = np.argmax(scores)
|
292 |
+
|
293 |
+
final_mask = masks[best_idx]
|
294 |
+
mask_colors = np.zeros((final_mask.shape[0], final_mask.shape[1], 3), dtype=np.uint8)
|
295 |
+
mask_colors[final_mask, :] = np.array([[128, 0, 0]])
|
296 |
+
output_image.append(Image.fromarray((mask_colors * 0.6 + test_image * 0.4).astype('uint8'), 'RGB'))
|
297 |
+
|
298 |
+
return output_image[0].resize((224, 224)), output_image[1].resize((224, 224))
|
299 |
+
|
300 |
+
|
301 |
+
def inference_finetune(ic_image, ic_mask, image1, image2):
|
302 |
+
# in context image and mask
|
303 |
+
ic_image = np.array(ic_image.convert("RGB"))
|
304 |
+
ic_mask = np.array(ic_mask.convert("RGB"))
|
305 |
+
|
306 |
+
gt_mask = torch.tensor(ic_mask)[:, :, 0] > 0
|
307 |
+
gt_mask = gt_mask.float().unsqueeze(0).flatten(1).cuda()
|
308 |
+
# gt_mask = gt_mask.float().unsqueeze(0).flatten(1)
|
309 |
+
|
310 |
+
sam_type, sam_ckpt = 'vit_h', 'sam_vit_h_4b8939.pth'
|
311 |
+
sam = sam_model_registry[sam_type](checkpoint=sam_ckpt).cuda()
|
312 |
+
# sam = sam_model_registry[sam_type](checkpoint=sam_ckpt)
|
313 |
+
for name, param in sam.named_parameters():
|
314 |
+
param.requires_grad = False
|
315 |
+
predictor = SamPredictor(sam)
|
316 |
+
|
317 |
+
print("======> Obtain Self Location Prior" )
|
318 |
+
# Image features encoding
|
319 |
+
ref_mask = predictor.set_image(ic_image, ic_mask)
|
320 |
+
ref_feat = predictor.features.squeeze().permute(1, 2, 0)
|
321 |
+
|
322 |
+
ref_mask = F.interpolate(ref_mask, size=ref_feat.shape[0: 2], mode="bilinear")
|
323 |
+
ref_mask = ref_mask.squeeze()[0]
|
324 |
+
|
325 |
+
# Target feature extraction
|
326 |
+
target_feat = ref_feat[ref_mask > 0]
|
327 |
+
target_feat_mean = target_feat.mean(0)
|
328 |
+
target_feat_max = torch.max(target_feat, dim=0)[0]
|
329 |
+
target_feat = (target_feat_max / 2 + target_feat_mean / 2).unsqueeze(0)
|
330 |
+
|
331 |
+
# Cosine similarity
|
332 |
+
h, w, C = ref_feat.shape
|
333 |
+
target_feat = target_feat / target_feat.norm(dim=-1, keepdim=True)
|
334 |
+
ref_feat = ref_feat / ref_feat.norm(dim=-1, keepdim=True)
|
335 |
+
ref_feat = ref_feat.permute(2, 0, 1).reshape(C, h * w)
|
336 |
+
sim = target_feat @ ref_feat
|
337 |
+
|
338 |
+
sim = sim.reshape(1, 1, h, w)
|
339 |
+
sim = F.interpolate(sim, scale_factor=4, mode="bilinear")
|
340 |
+
sim = predictor.model.postprocess_masks(
|
341 |
+
sim,
|
342 |
+
input_size=predictor.input_size,
|
343 |
+
original_size=predictor.original_size).squeeze()
|
344 |
+
|
345 |
+
# Positive location prior
|
346 |
+
topk_xy, topk_label, _, _ = point_selection(sim, topk=1)
|
347 |
+
|
348 |
+
print('======> Start Training')
|
349 |
+
# Learnable mask weights
|
350 |
+
mask_weights = Mask_Weights().cuda()
|
351 |
+
# mask_weights = Mask_Weights()
|
352 |
+
mask_weights.train()
|
353 |
+
train_epoch = 1000
|
354 |
+
optimizer = torch.optim.AdamW(mask_weights.parameters(), lr=1e-3, eps=1e-4)
|
355 |
+
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, train_epoch)
|
356 |
+
|
357 |
+
for train_idx in range(train_epoch):
|
358 |
+
# Run the decoder
|
359 |
+
masks, scores, logits, logits_high = predictor.predict(
|
360 |
+
point_coords=topk_xy,
|
361 |
+
point_labels=topk_label,
|
362 |
+
multimask_output=True)
|
363 |
+
logits_high = logits_high.flatten(1)
|
364 |
+
|
365 |
+
# Weighted sum three-scale masks
|
366 |
+
weights = torch.cat((1 - mask_weights.weights.sum(0).unsqueeze(0), mask_weights.weights), dim=0)
|
367 |
+
logits_high = logits_high * weights
|
368 |
+
logits_high = logits_high.sum(0).unsqueeze(0)
|
369 |
+
|
370 |
+
dice_loss = calculate_dice_loss(logits_high, gt_mask)
|
371 |
+
focal_loss = calculate_sigmoid_focal_loss(logits_high, gt_mask)
|
372 |
+
loss = dice_loss + focal_loss
|
373 |
+
|
374 |
+
optimizer.zero_grad()
|
375 |
+
loss.backward()
|
376 |
+
optimizer.step()
|
377 |
+
scheduler.step()
|
378 |
+
|
379 |
+
if train_idx % 10 == 0:
|
380 |
+
print('Train Epoch: {:} / {:}'.format(train_idx, train_epoch))
|
381 |
+
current_lr = scheduler.get_last_lr()[0]
|
382 |
+
print('LR: {:.6f}, Dice_Loss: {:.4f}, Focal_Loss: {:.4f}'.format(current_lr, dice_loss.item(), focal_loss.item()))
|
383 |
+
|
384 |
+
|
385 |
+
mask_weights.eval()
|
386 |
+
weights = torch.cat((1 - mask_weights.weights.sum(0).unsqueeze(0), mask_weights.weights), dim=0)
|
387 |
+
weights_np = weights.detach().cpu().numpy()
|
388 |
+
print('======> Mask weights:\n', weights_np)
|
389 |
+
|
390 |
+
print('======> Start Testing')
|
391 |
+
output_image = []
|
392 |
+
|
393 |
+
for test_image in [image1, image2]:
|
394 |
+
test_image = np.array(test_image.convert("RGB"))
|
395 |
+
|
396 |
+
# Image feature encoding
|
397 |
+
predictor.set_image(test_image)
|
398 |
+
test_feat = predictor.features.squeeze()
|
399 |
+
# Image feature encoding
|
400 |
+
predictor.set_image(test_image)
|
401 |
+
test_feat = predictor.features.squeeze()
|
402 |
+
|
403 |
+
# Cosine similarity
|
404 |
+
C, h, w = test_feat.shape
|
405 |
+
test_feat = test_feat / test_feat.norm(dim=0, keepdim=True)
|
406 |
+
test_feat = test_feat.reshape(C, h * w)
|
407 |
+
sim = target_feat @ test_feat
|
408 |
+
|
409 |
+
sim = sim.reshape(1, 1, h, w)
|
410 |
+
sim = F.interpolate(sim, scale_factor=4, mode="bilinear")
|
411 |
+
sim = predictor.model.postprocess_masks(
|
412 |
+
sim,
|
413 |
+
input_size=predictor.input_size,
|
414 |
+
original_size=predictor.original_size).squeeze()
|
415 |
+
|
416 |
+
# Positive location prior
|
417 |
+
topk_xy, topk_label, _, _ = point_selection(sim, topk=1)
|
418 |
+
|
419 |
+
# First-step prediction
|
420 |
+
masks, scores, logits, logits_high = predictor.predict(
|
421 |
+
point_coords=topk_xy,
|
422 |
+
point_labels=topk_label,
|
423 |
+
multimask_output=True)
|
424 |
+
|
425 |
+
# Weighted sum three-scale masks
|
426 |
+
logits_high = logits_high * weights.unsqueeze(-1)
|
427 |
+
logit_high = logits_high.sum(0)
|
428 |
+
mask = (logit_high > 0).detach().cpu().numpy()
|
429 |
+
|
430 |
+
logits = logits * weights_np[..., None]
|
431 |
+
logit = logits.sum(0)
|
432 |
+
|
433 |
+
# Cascaded Post-refinement-1
|
434 |
+
y, x = np.nonzero(mask)
|
435 |
+
x_min = x.min()
|
436 |
+
x_max = x.max()
|
437 |
+
y_min = y.min()
|
438 |
+
y_max = y.max()
|
439 |
+
input_box = np.array([x_min, y_min, x_max, y_max])
|
440 |
+
masks, scores, logits, _ = predictor.predict(
|
441 |
+
point_coords=topk_xy,
|
442 |
+
point_labels=topk_label,
|
443 |
+
box=input_box[None, :],
|
444 |
+
mask_input=logit[None, :, :],
|
445 |
+
multimask_output=True)
|
446 |
+
best_idx = np.argmax(scores)
|
447 |
+
|
448 |
+
# Cascaded Post-refinement-2
|
449 |
+
y, x = np.nonzero(masks[best_idx])
|
450 |
+
x_min = x.min()
|
451 |
+
x_max = x.max()
|
452 |
+
y_min = y.min()
|
453 |
+
y_max = y.max()
|
454 |
+
input_box = np.array([x_min, y_min, x_max, y_max])
|
455 |
+
masks, scores, logits, _ = predictor.predict(
|
456 |
+
point_coords=topk_xy,
|
457 |
+
point_labels=topk_label,
|
458 |
+
box=input_box[None, :],
|
459 |
+
mask_input=logits[best_idx: best_idx + 1, :, :],
|
460 |
+
multimask_output=True)
|
461 |
+
best_idx = np.argmax(scores)
|
462 |
+
|
463 |
+
final_mask = masks[best_idx]
|
464 |
+
mask_colors = np.zeros((final_mask.shape[0], final_mask.shape[1], 3), dtype=np.uint8)
|
465 |
+
mask_colors[final_mask, :] = np.array([[128, 0, 0]])
|
466 |
+
output_image.append(Image.fromarray((mask_colors * 0.6 + test_image * 0.4).astype('uint8'), 'RGB'))
|
467 |
+
|
468 |
+
return output_image[0].resize((224, 224)), output_image[1].resize((224, 224))
|
469 |
+
|
470 |
+
|
471 |
+
description = """
|
472 |
+
<div style="text-align: center; font-weight: bold;">
|
473 |
+
<span style="font-size: 18px" id="paper-info">
|
474 |
+
[<a href="https://github.com/ZrrSkywalker/Personalize-SAM" target="_blank">GitHub</a>]
|
475 |
+
[<a href="https://arxiv.org/pdf/2305.03048.pdf" target="_blank">Paper</a>]
|
476 |
+
</span>
|
477 |
+
</div>
|
478 |
+
"""
|
479 |
+
|
480 |
+
main = gr.Interface(
|
481 |
+
fn=inference,
|
482 |
+
inputs=[
|
483 |
+
gr.Image(type="pil", label="in context image",),
|
484 |
+
gr.Image(type="pil", label="in context mask"),
|
485 |
+
gr.Image(type="pil", label="test image1"),
|
486 |
+
gr.Image(type="pil", label="test image2"),
|
487 |
+
],
|
488 |
+
outputs=[
|
489 |
+
gr.outputs.Image(type="pil", label="output image1"),
|
490 |
+
gr.outputs.Image(type="pil", label="output image2"),
|
491 |
+
],
|
492 |
+
allow_flagging="never",
|
493 |
+
title="Personalize Segment Anything Model with 1 Shot",
|
494 |
+
description=description,
|
495 |
+
examples=[
|
496 |
+
["./examples/cat_00.jpg", "./examples/cat_00.png", "./examples/cat_01.jpg", "./examples/cat_02.jpg"],
|
497 |
+
["./examples/colorful_sneaker_00.jpg", "./examples/colorful_sneaker_00.png", "./examples/colorful_sneaker_01.jpg", "./examples/colorful_sneaker_02.jpg"],
|
498 |
+
["./examples/duck_toy_00.jpg", "./examples/duck_toy_00.png", "./examples/duck_toy_01.jpg", "./examples/duck_toy_02.jpg"],
|
499 |
+
]
|
500 |
+
)
|
501 |
+
|
502 |
+
main_scribble = gr.Interface(
|
503 |
+
fn=inference_scribble,
|
504 |
+
inputs=[
|
505 |
+
gr.ImageMask(label="[Stroke] Draw on Image", type="pil"),
|
506 |
+
gr.Image(type="pil", label="test image1"),
|
507 |
+
gr.Image(type="pil", label="test image2"),
|
508 |
+
],
|
509 |
+
outputs=[
|
510 |
+
gr.outputs.Image(type="pil", label="output image1"),
|
511 |
+
gr.outputs.Image(type="pil", label="output image2"),
|
512 |
+
],
|
513 |
+
allow_flagging="never",
|
514 |
+
title="Personalize Segment Anything Model with 1 Shot",
|
515 |
+
description=description,
|
516 |
+
examples=[
|
517 |
+
["./examples/cat_00.jpg", "./examples/cat_01.jpg", "./examples/cat_02.jpg"],
|
518 |
+
["./examples/colorful_sneaker_00.jpg", "./examples/colorful_sneaker_01.jpg", "./examples/colorful_sneaker_02.jpg"],
|
519 |
+
["./examples/duck_toy_00.jpg", "./examples/duck_toy_01.jpg", "./examples/duck_toy_02.jpg"],
|
520 |
+
]
|
521 |
+
)
|
522 |
+
|
523 |
+
main_finetune = gr.Interface(
|
524 |
+
fn=inference_finetune,
|
525 |
+
inputs=[
|
526 |
+
gr.Image(type="pil", label="in context image"),
|
527 |
+
gr.Image(type="pil", label="in context mask"),
|
528 |
+
gr.Image(type="pil", label="test image1"),
|
529 |
+
gr.Image(type="pil", label="test image2"),
|
530 |
+
],
|
531 |
+
outputs=[
|
532 |
+
gr.components.Image(type="pil", label="output image1"),
|
533 |
+
gr.components.Image(type="pil", label="output image2"),
|
534 |
+
],
|
535 |
+
allow_flagging="never",
|
536 |
+
title="Personalize Segment Anything Model with 1 Shot",
|
537 |
+
description=description,
|
538 |
+
examples=[
|
539 |
+
["./examples/cat_00.jpg", "./examples/cat_00.png", "./examples/cat_01.jpg", "./examples/cat_02.jpg"],
|
540 |
+
["./examples/colorful_sneaker_00.jpg", "./examples/colorful_sneaker_00.png", "./examples/colorful_sneaker_01.jpg", "./examples/colorful_sneaker_02.jpg"],
|
541 |
+
["./examples/duck_toy_00.jpg", "./examples/duck_toy_00.png", "./examples/duck_toy_01.jpg", "./examples/duck_toy_02.jpg"],
|
542 |
+
]
|
543 |
+
)
|
544 |
+
|
545 |
+
|
546 |
+
demo = gr.Blocks()
|
547 |
+
with demo:
|
548 |
+
gr.TabbedInterface(
|
549 |
+
[main, main_scribble, main_finetune],
|
550 |
+
["Personalize-SAM", "Personalize-SAM-Scribble", "Personalize-SAM-F"],
|
551 |
+
)
|
552 |
+
|
553 |
+
demo.launch(share=True)
|
requirements.txt
ADDED
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
matplotlib
|
2 |
+
tqdm
|
3 |
+
os
|
4 |
+
numpy
|
5 |
+
warnings
|
6 |
+
argparse
|
7 |
+
opencv-python
|
show.py
ADDED
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
import torch
|
3 |
+
import matplotlib.pyplot as plt
|
4 |
+
import cv2
|
5 |
+
|
6 |
+
|
7 |
+
|
8 |
+
def show_mask(mask, ax, random_color=False):
|
9 |
+
if random_color:
|
10 |
+
color = np.concatenate([np.random.random(3), np.array([0.6])], axis=0)
|
11 |
+
else:
|
12 |
+
color = np.array([30/255, 144/255, 255/255, 0.4])
|
13 |
+
h, w = mask.shape[-2:]
|
14 |
+
mask_image = mask.reshape(h, w, 1) * color.reshape(1, 1, -1)
|
15 |
+
ax.imshow(mask_image)
|
16 |
+
|
17 |
+
|
18 |
+
def show_points(coords, labels, ax, marker_size=375):
|
19 |
+
pos_points = coords[labels==1]
|
20 |
+
neg_points = coords[labels==0]
|
21 |
+
ax.scatter(pos_points[:, 0], pos_points[:, 1], color='green', marker='*', s=marker_size, edgecolor='white', linewidth=1.25)
|
22 |
+
ax.scatter(neg_points[:, 0], neg_points[:, 1], color='red', marker='*', s=marker_size, edgecolor='white', linewidth=1.25)
|
23 |
+
|
24 |
+
|
25 |
+
def show_box(box, ax):
|
26 |
+
x0, y0 = box[0], box[1]
|
27 |
+
w, h = box[2] - box[0], box[3] - box[1]
|
28 |
+
ax.add_patch(plt.Rectangle((x0, y0), w, h, edgecolor='green', facecolor=(0,0,0,0), lw=2))
|